Unsupported-syntax diagnostics are preempted by closure-variable resolution

[egparedes]

Constructs that gt4py.next rejects via the _UNSUPPORTED_FEATURE_HINTS catalogue in ffront/dialect_parser.py only produce their friendly diagnostic when they do not reference an external name. If the construct names a symbol (e.g. an exception type), closure-variable resolution runs first and fails with an unrelated, internal-sounding message.

Reproduction

import gt4py.next as gtx
from gt4py.next import float64
from gt4py.next.ffront.func_to_foast import FieldOperatorParser

IDim = gtx.Dimension("IDim")


def try_finally(a: gtx.Field[[IDim], float64]) -> gtx.Field[[IDim], float64]:
    try:
        a = a + 1.0
    finally:
        pass
    return a


def try_except(a: gtx.Field[[IDim], float64]) -> gtx.Field[[IDim], float64]:
    try:
        a = a + 1.0
    except ValueError:
        pass
    return a

Results:

program	diagnostic
try_finally	UnsupportedPythonFeatureError: Unsupported Python syntax: 'try' statement. (correct)
try_except	DSLTypeError: Unexpected object 'ValueError' of type '<class 'type'>' encountered. (wrong)

Why

ValueError is collected as a closure variable and converted to a FOAST symbol before the AST is visited, so DialectParser.generic_visit — which consults the catalogue — is never reached.

Impact

• try / except X is the common spelling; the catalogued ast.Try message is effectively only reachable via the rare try / finally form.
• ast.TryStar is entirely unreachable: try* grammatically requires an except* handler, which must name an exception type. The catalogue entry added in fix[cartesian, eve, next]: latent bugs and deprecated typing shims surfaced by the Python 3.12 floor #2755 is therefore pinned at the catalogue level rather than end-to-end, with a comment explaining why.
• ast.Raise is likely affected in the same way (raise ValueError(...)).

The user-visible effect is that a plain unsupported-syntax error is reported as an internal type error with no actionable hint.

Possible direction

Visit the AST for unsupported constructs before (or independently of) resolving closure variables, so syntax-level rejections win over symbol-level ones. Adding a test for the try / except X spelling in tests/next_tests/unit_tests/ffront_tests/test_diagnostic_messages.py would pin the fix.

Found while removing the Python 3.10 shims in #2755.